Spaces:
Runtime error
Runtime error
Andrey
commited on
Commit
•
b683920
1
Parent(s):
5182668
Initial commit.
Browse files- .flake8 +5 -0
- .gitattributes +1 -0
- .idea/.gitignore +8 -0
- .pre-commit-config.yaml +35 -0
- README.md +15 -0
- config.toml +8 -0
- model_files/best_model.pth +3 -0
- mypy.ini +5 -0
- pyproject.toml +20 -0
- requirements.txt +15 -0
- src/ml_utils.py +207 -0
- src/model_architecture.py +156 -0
- src/utils.py +61 -0
- st_app.py +46 -0
.flake8
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[flake8]
|
2 |
+
ignore = I001,I002,I004,I005,I101,I201,C101,C403,C901,F401,F403,S001,D100,D101,D102,D103,D104,D105,D106,D107,D200,D210,D205,D400,T001,W504,D202,E203,W503,B006,T002,T100,P103,C408,F841
|
3 |
+
max-line-length = 120
|
4 |
+
exclude = outputs/*
|
5 |
+
max-complexity = 10
|
.gitattributes
CHANGED
@@ -1,2 +1,3 @@
|
|
1 |
# Auto detect text files and perform LF normalization
|
2 |
* text=auto
|
|
|
|
1 |
# Auto detect text files and perform LF normalization
|
2 |
* text=auto
|
3 |
+
*.pth filter=lfs diff=lfs merge=lfs -text
|
.idea/.gitignore
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Default ignored files
|
2 |
+
/shelf/
|
3 |
+
/workspace.xml
|
4 |
+
# Editor-based HTTP Client requests
|
5 |
+
/httpRequests/
|
6 |
+
# Datasource local storage ignored files
|
7 |
+
/dataSources/
|
8 |
+
/dataSources.local.xml
|
.pre-commit-config.yaml
ADDED
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
repos:
|
2 |
+
- repo: https://github.com/pre-commit/pre-commit-hooks
|
3 |
+
rev: v4.3.0
|
4 |
+
hooks:
|
5 |
+
- id: check-yaml
|
6 |
+
- id: end-of-file-fixer
|
7 |
+
- id: trailing-whitespace
|
8 |
+
- repo: https://github.com/psf/black
|
9 |
+
rev: '22.12.0'
|
10 |
+
hooks:
|
11 |
+
- id: black
|
12 |
+
args: [--config=pyproject.toml]
|
13 |
+
- repo: https://github.com/pre-commit/mirrors-mypy
|
14 |
+
rev: 586b4f0
|
15 |
+
hooks:
|
16 |
+
- id: mypy
|
17 |
+
args: [--ignore-missing-imports, --warn-no-return, --warn-redundant-casts, --disallow-incomplete-defs, --no-namespace-packages ]
|
18 |
+
- repo: https://gitlab.com/pycqa/flake8
|
19 |
+
rev: '5.0.4'
|
20 |
+
hooks:
|
21 |
+
- id: flake8
|
22 |
+
additional_dependencies: [
|
23 |
+
'flake8-bugbear==22.8.23',
|
24 |
+
'flake8-coding==1.3.2',
|
25 |
+
'flake8-comprehensions==3.10.0',
|
26 |
+
'flake8-debugger==4.1.2',
|
27 |
+
'flake8-deprecated==1.3',
|
28 |
+
'flake8-docstrings==1.6.0',
|
29 |
+
'flake8-isort==4.2.0',
|
30 |
+
'flake8-pep3101==1.3.0',
|
31 |
+
'flake8-polyfill==1.0.2',
|
32 |
+
'flake8-print==5.0.0',
|
33 |
+
'flake8-quotes==3.3.1',
|
34 |
+
'flake8-string-format==0.3.0',
|
35 |
+
]
|
README.md
CHANGED
@@ -1,2 +1,17 @@
|
|
1 |
# digit-draw-detect
|
2 |
An app for handwritten digit detection
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
# digit-draw-detect
|
2 |
An app for handwritten digit detection
|
3 |
+
|
4 |
+
steps:
|
5 |
+
* use git lfs for the model +
|
6 |
+
* write better code +
|
7 |
+
* convert model to onnx or some other format?
|
8 |
+
* deploy bare working app, without nice things
|
9 |
+
* make better design
|
10 |
+
* think about descriptions on the site
|
11 |
+
|
12 |
+
On using git lfs:
|
13 |
+
```shell
|
14 |
+
git lfs install
|
15 |
+
git lfs track "*.psd"
|
16 |
+
git add .gitattributes
|
17 |
+
```
|
config.toml
ADDED
@@ -0,0 +1,8 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Everything in this section will be available as an environment variable
|
2 |
+
db_username = "Jane"
|
3 |
+
db_password = "12345qwerty"
|
4 |
+
|
5 |
+
AWS_ACCESS_KEY_ID = 'AKIAI4JDKBYRCHGT77VQ'
|
6 |
+
AWS_SECRET_ACCESS_KEY = 'ewSheQRxUKM/QTtHUPlESpMhl4bBQfihGWpBFy4s'
|
7 |
+
S3_BUCKET = 'digitdrawdetect'
|
8 |
+
S3_BUCKET_NAME = 'digitdrawdetect'
|
model_files/best_model.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:571e937122d5ccafe496d1cc71cea5c0661d385b5a7db4ec977ac8ae5da40680
|
3 |
+
size 246698572
|
mypy.ini
ADDED
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy
|
2 |
+
|
3 |
+
[mypy]
|
4 |
+
python_version = 3.10
|
5 |
+
plugins = numpy.typing.mypy_plugin
|
pyproject.toml
ADDED
@@ -0,0 +1,20 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
[tool.black]
|
2 |
+
line-length = 119
|
3 |
+
skip-string-normalization = true
|
4 |
+
target-version = ['py39', 'py310']
|
5 |
+
include = '\.pyi?$'
|
6 |
+
exclude = '''
|
7 |
+
/(
|
8 |
+
\.eggs
|
9 |
+
| \.git
|
10 |
+
| \.hg
|
11 |
+
| \.mypy_cache
|
12 |
+
| \.tox
|
13 |
+
| \.venv
|
14 |
+
| _build
|
15 |
+
| buck-out
|
16 |
+
| build
|
17 |
+
| dist
|
18 |
+
| outputs
|
19 |
+
)/
|
20 |
+
'''
|
requirements.txt
ADDED
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
albumentations==1.3.0
|
2 |
+
matplotlib==3.6.1
|
3 |
+
numpy==1.23.4
|
4 |
+
omegaconf==2.2.1
|
5 |
+
opencv_python==4.6.0.66
|
6 |
+
pandas==1.5.1
|
7 |
+
Pillow==9.2.0
|
8 |
+
rich==12.6.0
|
9 |
+
streamlit==1.16.0
|
10 |
+
streamlit_drawable_canvas==0.9.2
|
11 |
+
tomli==2.0.1
|
12 |
+
torch==1.12.1
|
13 |
+
torchvision
|
14 |
+
# need to define some pytorch
|
15 |
+
https://download.pytorch.org/whl/cpu/torch-1.13.1%2Bcpu-cp310-cp310-linux_x86_64.whl
|
src/ml_utils.py
ADDED
@@ -0,0 +1,207 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List
|
2 |
+
import logging
|
3 |
+
|
4 |
+
import albumentations as A
|
5 |
+
import streamlit as st
|
6 |
+
import torch
|
7 |
+
from albumentations import pytorch
|
8 |
+
|
9 |
+
from src.model_architecture import Net
|
10 |
+
|
11 |
+
anchors = torch.tensor(
|
12 |
+
[
|
13 |
+
[[0.2800, 0.2200], [0.3800, 0.4800], [0.9000, 0.7800]],
|
14 |
+
[[0.0700, 0.1500], [0.1500, 0.1100], [0.1400, 0.2900]],
|
15 |
+
[[0.0200, 0.0300], [0.0400, 0.0700], [0.0800, 0.0600]],
|
16 |
+
]
|
17 |
+
)
|
18 |
+
|
19 |
+
transforms = A.Compose(
|
20 |
+
[
|
21 |
+
A.Resize(always_apply=False, p=1, height=192, width=192, interpolation=1),
|
22 |
+
A.Normalize(),
|
23 |
+
pytorch.transforms.ToTensorV2(),
|
24 |
+
]
|
25 |
+
)
|
26 |
+
|
27 |
+
|
28 |
+
def cells_to_bboxes(predictions: torch.tensor, anchors: torch.tensor, s: int, is_preds: bool = True) -> torch.tensor:
|
29 |
+
"""
|
30 |
+
Scale the predictions coming from the model_files to
|
31 |
+
be relative to the entire image such that they for example later
|
32 |
+
can be plotted or.
|
33 |
+
Args:
|
34 |
+
predictions: tensor of size (N, 3, S, S, num_classes+5)
|
35 |
+
anchors: the anchors used for the predictions
|
36 |
+
s: the number of cells the image is divided in on the width (and height)
|
37 |
+
is_preds: whether the input is predictions or the true bounding boxes
|
38 |
+
Returns:
|
39 |
+
converted_bboxes: the converted boxes of sizes (N, num_anchors, S, S, 1+5) with class index,
|
40 |
+
object score, bounding box coordinates
|
41 |
+
"""
|
42 |
+
batch_size = predictions.shape[0]
|
43 |
+
num_anchors = len(anchors)
|
44 |
+
box_predictions = predictions[..., 1:5]
|
45 |
+
if is_preds:
|
46 |
+
anchors = anchors.reshape(1, len(anchors), 1, 1, 2)
|
47 |
+
box_predictions[..., 0:2] = torch.sigmoid(box_predictions[..., 0:2])
|
48 |
+
box_predictions[..., 2:] = torch.exp(box_predictions[..., 2:]) * anchors
|
49 |
+
scores = torch.sigmoid(predictions[..., 0:1])
|
50 |
+
best_class = torch.argmax(predictions[..., 5:], dim=-1).unsqueeze(-1)
|
51 |
+
else:
|
52 |
+
scores = predictions[..., 0:1]
|
53 |
+
best_class = predictions[..., 5:6]
|
54 |
+
|
55 |
+
cell_indices = torch.arange(s).repeat(predictions.shape[0], 3, s, 1).unsqueeze(-1).to(predictions.device)
|
56 |
+
x = 1 / s * (box_predictions[..., 0:1] + cell_indices)
|
57 |
+
y = 1 / s * (box_predictions[..., 1:2] + cell_indices.permute(0, 1, 3, 2, 4))
|
58 |
+
w_h = 1 / s * box_predictions[..., 2:4]
|
59 |
+
converted_bboxes = torch.cat((best_class, scores, x, y, w_h), dim=-1).reshape(batch_size, num_anchors * s * s, 6)
|
60 |
+
return converted_bboxes.tolist()
|
61 |
+
|
62 |
+
|
63 |
+
def non_max_suppression(
|
64 |
+
bboxes: List[List], iou_threshold: float, threshold: float, box_format: str = 'corners'
|
65 |
+
) -> List[List]:
|
66 |
+
"""
|
67 |
+
Apply nms to the bboxes.
|
68 |
+
|
69 |
+
Video explanation of this function:
|
70 |
+
https://youtu.be/YDkjWEN8jNA
|
71 |
+
Does Non Max Suppression given bboxes
|
72 |
+
Args:
|
73 |
+
bboxes (list): list of lists containing all bboxes with each bboxes
|
74 |
+
specified as [class_pred, prob_score, x1, y1, x2, y2]
|
75 |
+
iou_threshold (float): threshold where predicted bboxes is correct
|
76 |
+
threshold (float): threshold to remove predicted bboxes (independent of IoU)
|
77 |
+
box_format (str): 'midpoint' or 'corners' used to specify bboxes
|
78 |
+
Returns:
|
79 |
+
list: bboxes after performing NMS given a specific IoU threshold
|
80 |
+
"""
|
81 |
+
|
82 |
+
assert type(bboxes) == list
|
83 |
+
|
84 |
+
bboxes = [box for box in bboxes if box[1] > threshold]
|
85 |
+
bboxes = sorted(bboxes, key=lambda x: x[1], reverse=True)
|
86 |
+
bboxes_after_nms = []
|
87 |
+
|
88 |
+
while bboxes:
|
89 |
+
chosen_box = bboxes.pop(0)
|
90 |
+
|
91 |
+
bboxes = [
|
92 |
+
box
|
93 |
+
for box in bboxes
|
94 |
+
if box[0] != chosen_box[0]
|
95 |
+
or intersection_over_union(
|
96 |
+
torch.tensor(chosen_box[2:]),
|
97 |
+
torch.tensor(box[2:]),
|
98 |
+
box_format=box_format,
|
99 |
+
)
|
100 |
+
< iou_threshold
|
101 |
+
]
|
102 |
+
|
103 |
+
bboxes_after_nms.append(chosen_box)
|
104 |
+
|
105 |
+
return bboxes_after_nms
|
106 |
+
|
107 |
+
|
108 |
+
def intersection_over_union(
|
109 |
+
boxes_preds: torch.tensor, boxes_labels: torch.tensor, box_format: str = 'midpoint'
|
110 |
+
) -> torch.tensor:
|
111 |
+
"""
|
112 |
+
Calculate iou.
|
113 |
+
|
114 |
+
Video explanation of this function:
|
115 |
+
https://youtu.be/XXYG5ZWtjj0
|
116 |
+
This function calculates intersection over union (iou) given pred boxes
|
117 |
+
and target boxes.
|
118 |
+
Args:
|
119 |
+
boxes_preds (tensor): Predictions of Bounding Boxes (BATCH_SIZE, 4)
|
120 |
+
boxes_labels (tensor): Correct labels of Bounding Boxes (BATCH_SIZE, 4)
|
121 |
+
box_format (str): midpoint/corners, if boxes (x,y,w,h) or (x1,y1,x2,y2)
|
122 |
+
Returns:
|
123 |
+
tensor: Intersection over union for all examples
|
124 |
+
"""
|
125 |
+
|
126 |
+
if box_format == 'midpoint':
|
127 |
+
box1_x1 = boxes_preds[..., 0:1] - boxes_preds[..., 2:3] / 2
|
128 |
+
box1_y1 = boxes_preds[..., 1:2] - boxes_preds[..., 3:4] / 2
|
129 |
+
box1_x2 = boxes_preds[..., 0:1] + boxes_preds[..., 2:3] / 2
|
130 |
+
box1_y2 = boxes_preds[..., 1:2] + boxes_preds[..., 3:4] / 2
|
131 |
+
box2_x1 = boxes_labels[..., 0:1] - boxes_labels[..., 2:3] / 2
|
132 |
+
box2_y1 = boxes_labels[..., 1:2] - boxes_labels[..., 3:4] / 2
|
133 |
+
box2_x2 = boxes_labels[..., 0:1] + boxes_labels[..., 2:3] / 2
|
134 |
+
box2_y2 = boxes_labels[..., 1:2] + boxes_labels[..., 3:4] / 2
|
135 |
+
|
136 |
+
if box_format == 'corners':
|
137 |
+
box1_x1 = boxes_preds[..., 0:1]
|
138 |
+
box1_y1 = boxes_preds[..., 1:2]
|
139 |
+
box1_x2 = boxes_preds[..., 2:3]
|
140 |
+
box1_y2 = boxes_preds[..., 3:4]
|
141 |
+
box2_x1 = boxes_labels[..., 0:1]
|
142 |
+
box2_y1 = boxes_labels[..., 1:2]
|
143 |
+
box2_x2 = boxes_labels[..., 2:3]
|
144 |
+
box2_y2 = boxes_labels[..., 3:4]
|
145 |
+
|
146 |
+
x1 = torch.max(box1_x1, box2_x1)
|
147 |
+
y1 = torch.max(box1_y1, box2_y1)
|
148 |
+
x2 = torch.min(box1_x2, box2_x2)
|
149 |
+
y2 = torch.min(box1_y2, box2_y2)
|
150 |
+
|
151 |
+
intersection = (x2 - x1).clamp(0) * (y2 - y1).clamp(0)
|
152 |
+
box1_area = abs((box1_x2 - box1_x1) * (box1_y2 - box1_y1))
|
153 |
+
box2_area = abs((box2_x2 - box2_x1) * (box2_y2 - box2_y1))
|
154 |
+
|
155 |
+
return intersection / (box1_area + box2_area - intersection + 1e-6)
|
156 |
+
|
157 |
+
|
158 |
+
def predict(
|
159 |
+
model: torch.nn.Module, image: torch.tensor, iou_threshold: float = 1.0, threshold: float = 0.05
|
160 |
+
) -> List[List]:
|
161 |
+
"""
|
162 |
+
Apply the model_files to the predictions and to postprocessing
|
163 |
+
Args:
|
164 |
+
model: a trained pytorch model_files.
|
165 |
+
image: image as a torch tensor
|
166 |
+
iou_threshold: a threshold for intersection_over_union function
|
167 |
+
threshold: a threshold for bbox probability
|
168 |
+
|
169 |
+
Returns:
|
170 |
+
predicted bboxes
|
171 |
+
|
172 |
+
"""
|
173 |
+
# apply model_files. add a dimension to imitate a batch size of 1
|
174 |
+
logits = model(image[None, :])
|
175 |
+
logging.info('predicted')
|
176 |
+
|
177 |
+
# postprocess. In fact, we could remove indexing with idx here, as there is a single image.
|
178 |
+
# But I prefer to keep it so that this code could be easier changed for cases with batch size > 1
|
179 |
+
bboxes: List[List] = [[] for _ in range(1)]
|
180 |
+
for i in range(3):
|
181 |
+
S = logits[i].shape[2]
|
182 |
+
# it could be better to initialize anchors inside the function, but I don't want to do it for every prediction.
|
183 |
+
anchor = anchors[i] * S
|
184 |
+
boxes_scale_i = cells_to_bboxes(logits[i], anchor, s=S, is_preds=True)
|
185 |
+
for idx, (box) in enumerate(boxes_scale_i):
|
186 |
+
bboxes[idx] += box
|
187 |
+
logging.info('Starting nms')
|
188 |
+
nms_boxes = non_max_suppression(
|
189 |
+
bboxes[idx],
|
190 |
+
iou_threshold=iou_threshold,
|
191 |
+
threshold=threshold,
|
192 |
+
box_format='midpoint',
|
193 |
+
)
|
194 |
+
|
195 |
+
return nms_boxes
|
196 |
+
|
197 |
+
|
198 |
+
@st.cache
|
199 |
+
def get_model():
|
200 |
+
|
201 |
+
model_name = 'model_files/best_model.pth'
|
202 |
+
|
203 |
+
model = Net()
|
204 |
+
model.load_state_dict(torch.load(model_name))
|
205 |
+
model.eval()
|
206 |
+
|
207 |
+
return model
|
src/model_architecture.py
ADDED
@@ -0,0 +1,156 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
|
4 |
+
|
5 |
+
class CNNBlock(nn.Module):
|
6 |
+
def __init__(self, in_channels, out_channels, bn_act=True, **kwargs):
|
7 |
+
super().__init__()
|
8 |
+
self.conv = nn.Conv2d(in_channels, out_channels, bias=not bn_act, **kwargs)
|
9 |
+
self.bn = nn.BatchNorm2d(out_channels)
|
10 |
+
self.leaky = nn.LeakyReLU(0.1)
|
11 |
+
self.use_bn_act = bn_act
|
12 |
+
|
13 |
+
def forward(self, x):
|
14 |
+
if self.use_bn_act:
|
15 |
+
return self.leaky(self.bn(self.conv(x)))
|
16 |
+
else:
|
17 |
+
return self.conv(x)
|
18 |
+
|
19 |
+
|
20 |
+
class ResidualBlock(nn.Module):
|
21 |
+
def __init__(self, channels, use_residual=True, num_repeats=1):
|
22 |
+
super().__init__()
|
23 |
+
self.layers = nn.ModuleList()
|
24 |
+
for _ in range(num_repeats):
|
25 |
+
self.layers += [
|
26 |
+
nn.Sequential(
|
27 |
+
CNNBlock(channels, channels // 2, kernel_size=1),
|
28 |
+
CNNBlock(channels // 2, channels, kernel_size=3, padding=1),
|
29 |
+
)
|
30 |
+
]
|
31 |
+
|
32 |
+
self.use_residual = use_residual
|
33 |
+
self.num_repeats = num_repeats
|
34 |
+
|
35 |
+
def forward(self, x):
|
36 |
+
for layer in self.layers:
|
37 |
+
if self.use_residual:
|
38 |
+
x = x + layer(x)
|
39 |
+
else:
|
40 |
+
x = layer(x)
|
41 |
+
|
42 |
+
return x
|
43 |
+
|
44 |
+
|
45 |
+
class ScalePrediction(nn.Module):
|
46 |
+
def __init__(self, in_channels, num_classes):
|
47 |
+
super().__init__()
|
48 |
+
self.pred = nn.Sequential(
|
49 |
+
CNNBlock(in_channels, 2 * in_channels, kernel_size=3, padding=1),
|
50 |
+
CNNBlock(2 * in_channels, (num_classes + 5) * 3, bn_act=False, kernel_size=1),
|
51 |
+
)
|
52 |
+
self.num_classes = num_classes
|
53 |
+
|
54 |
+
def forward(self, x):
|
55 |
+
return self.pred(x).reshape(x.shape[0], 3, self.num_classes + 5, x.shape[2], x.shape[3]).permute(0, 1, 3, 4, 2)
|
56 |
+
|
57 |
+
|
58 |
+
class Net(nn.Module):
|
59 |
+
def __init__(self):
|
60 |
+
super().__init__()
|
61 |
+
self.num_classes = 12
|
62 |
+
self.in_channels = 3
|
63 |
+
# self.config = cfg.model_files.params.config
|
64 |
+
# self.config = [i if i[0] != '(' else literal_eval(i) for i in self.config]
|
65 |
+
self.config = [
|
66 |
+
(32, 3, 1),
|
67 |
+
(64, 3, 2),
|
68 |
+
['B', 1],
|
69 |
+
(128, 3, 2),
|
70 |
+
['B', 2],
|
71 |
+
(256, 3, 2),
|
72 |
+
['B', 8],
|
73 |
+
(512, 3, 2),
|
74 |
+
['B', 8],
|
75 |
+
(1024, 3, 2),
|
76 |
+
['B', 4],
|
77 |
+
(512, 1, 1),
|
78 |
+
(1024, 3, 1),
|
79 |
+
'S',
|
80 |
+
(256, 1, 1),
|
81 |
+
'U',
|
82 |
+
(256, 1, 1),
|
83 |
+
(512, 3, 1),
|
84 |
+
'S',
|
85 |
+
(128, 1, 1),
|
86 |
+
'U',
|
87 |
+
(128, 1, 1),
|
88 |
+
(256, 3, 1),
|
89 |
+
'S',
|
90 |
+
]
|
91 |
+
# print('self.config', self.config)
|
92 |
+
self.layers = self._create_conv_layers()
|
93 |
+
|
94 |
+
def forward(self, x):
|
95 |
+
outputs = [] # for each scale
|
96 |
+
route_connections = []
|
97 |
+
for layer in self.layers:
|
98 |
+
if isinstance(layer, ScalePrediction):
|
99 |
+
outputs.append(layer(x))
|
100 |
+
continue
|
101 |
+
# print(layer, x.shape)
|
102 |
+
x = layer(x)
|
103 |
+
|
104 |
+
if isinstance(layer, ResidualBlock) and layer.num_repeats == 8:
|
105 |
+
route_connections.append(x)
|
106 |
+
|
107 |
+
elif isinstance(layer, nn.Upsample):
|
108 |
+
x = torch.cat([x, route_connections[-1]], dim=1)
|
109 |
+
route_connections.pop()
|
110 |
+
|
111 |
+
return outputs
|
112 |
+
|
113 |
+
def _create_conv_layers(self):
|
114 |
+
layers = nn.ModuleList()
|
115 |
+
in_channels = self.in_channels
|
116 |
+
|
117 |
+
for module in self.config:
|
118 |
+
# print(module, type(module))
|
119 |
+
if isinstance(module, tuple):
|
120 |
+
out_channels, kernel_size, stride = module
|
121 |
+
layers.append(
|
122 |
+
CNNBlock(
|
123 |
+
in_channels,
|
124 |
+
out_channels,
|
125 |
+
kernel_size=kernel_size,
|
126 |
+
stride=stride,
|
127 |
+
padding=1 if kernel_size == 3 else 0,
|
128 |
+
)
|
129 |
+
)
|
130 |
+
in_channels = out_channels
|
131 |
+
|
132 |
+
elif isinstance(module, list):
|
133 |
+
num_repeats = module[1]
|
134 |
+
layers.append(
|
135 |
+
ResidualBlock(
|
136 |
+
in_channels,
|
137 |
+
num_repeats=num_repeats,
|
138 |
+
)
|
139 |
+
)
|
140 |
+
|
141 |
+
elif isinstance(module, str):
|
142 |
+
if module == 'S':
|
143 |
+
layers += [
|
144 |
+
ResidualBlock(in_channels, use_residual=False, num_repeats=1),
|
145 |
+
CNNBlock(in_channels, in_channels // 2, kernel_size=1),
|
146 |
+
ScalePrediction(in_channels // 2, num_classes=self.num_classes),
|
147 |
+
]
|
148 |
+
in_channels = in_channels // 2
|
149 |
+
|
150 |
+
elif module == 'U':
|
151 |
+
layers.append(
|
152 |
+
nn.Upsample(scale_factor=2),
|
153 |
+
)
|
154 |
+
in_channels = in_channels * 3
|
155 |
+
|
156 |
+
return layers
|
src/utils.py
ADDED
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import List, Dict
|
2 |
+
|
3 |
+
import matplotlib
|
4 |
+
import matplotlib.patches as patches
|
5 |
+
import matplotlib.pyplot as plt
|
6 |
+
import numpy as np
|
7 |
+
import tomli as tomllib
|
8 |
+
|
9 |
+
|
10 |
+
def plot_img_with_rects(
|
11 |
+
img: np.array, boxes: List[List], threshold: float = 0.5, coef: int = 400
|
12 |
+
) -> matplotlib.figure.Figure:
|
13 |
+
"""
|
14 |
+
Plot image with rectangles.
|
15 |
+
|
16 |
+
Args:
|
17 |
+
img: image as a numpy array
|
18 |
+
boxes: the list of the bboxes
|
19 |
+
threshold: threshold for bbox probability
|
20 |
+
coef: coefficient to multiply images. Can be changed when the original image is a different size
|
21 |
+
|
22 |
+
Returns:
|
23 |
+
image with bboxes
|
24 |
+
"""
|
25 |
+
fig, ax = plt.subplots(1, figsize=(4, 4))
|
26 |
+
|
27 |
+
# Display the image
|
28 |
+
ax.imshow(img)
|
29 |
+
|
30 |
+
# Create a Rectangle patch
|
31 |
+
for _, rect in enumerate([b for b in boxes if b[1] > threshold]):
|
32 |
+
label, _, xc, yc, w, h = rect
|
33 |
+
xc, yc, w, h = xc * coef, yc * coef, w * coef, h * coef
|
34 |
+
# the coordinates from center-based to left top corner
|
35 |
+
x = xc - w / 2
|
36 |
+
y = yc - h / 2
|
37 |
+
label = int(label)
|
38 |
+
label = label if label != 10 else 'penis'
|
39 |
+
label = label if label != 11 else 'junk'
|
40 |
+
rect = [x, y, x + w, y + h]
|
41 |
+
|
42 |
+
rect_ = patches.Rectangle(
|
43 |
+
(rect[0], rect[1]), rect[2] - rect[0], rect[3] - rect[1], linewidth=2, edgecolor='blue', facecolor='none'
|
44 |
+
)
|
45 |
+
plt.text(rect[2], rect[1], f'{label}', color='blue')
|
46 |
+
# Add the patch to the Axes
|
47 |
+
ax.add_patch(rect_)
|
48 |
+
return fig
|
49 |
+
|
50 |
+
|
51 |
+
def get_config() -> Dict:
|
52 |
+
"""
|
53 |
+
Get dict from config.
|
54 |
+
|
55 |
+
Returns:
|
56 |
+
config
|
57 |
+
"""
|
58 |
+
with open('config.toml', 'rb') as f:
|
59 |
+
config = tomllib.load(f)
|
60 |
+
|
61 |
+
return config
|
st_app.py
ADDED
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import logging
|
2 |
+
|
3 |
+
import numpy as np
|
4 |
+
import streamlit as st
|
5 |
+
import tomli as tomllib
|
6 |
+
from PIL import Image
|
7 |
+
from streamlit_drawable_canvas import st_canvas
|
8 |
+
|
9 |
+
from src.ml_utils import predict, get_model, transforms
|
10 |
+
from src.utils import plot_img_with_rects, get_config
|
11 |
+
|
12 |
+
logging.info('Starting')
|
13 |
+
|
14 |
+
col1, col2 = st.columns(2)
|
15 |
+
|
16 |
+
with col1:
|
17 |
+
# Create a canvas component
|
18 |
+
canvas_result = st_canvas(
|
19 |
+
fill_color='#fff',
|
20 |
+
stroke_width=5,
|
21 |
+
stroke_color='#000',
|
22 |
+
background_color='#fff',
|
23 |
+
update_streamlit=True,
|
24 |
+
height=400,
|
25 |
+
width=400,
|
26 |
+
drawing_mode='freedraw',
|
27 |
+
key='canvas',
|
28 |
+
)
|
29 |
+
with col2:
|
30 |
+
data = get_config()
|
31 |
+
logging.info('canvas ready')
|
32 |
+
if canvas_result.image_data is not None:
|
33 |
+
# convert a drawn image into numpy array with RGB from a canvas image with RGBA
|
34 |
+
img = np.array(Image.fromarray(np.uint8(canvas_result.image_data)).convert('RGB'))
|
35 |
+
image = transforms(image=img)['image']
|
36 |
+
logging.info('image augmented')
|
37 |
+
model = get_model()
|
38 |
+
logging.info('model ready')
|
39 |
+
pred = predict(model, image)
|
40 |
+
logging.info('prediction done')
|
41 |
+
threshold = st.slider('Bbox probability slider', min_value=0.0, max_value=1.0, value=0.5)
|
42 |
+
|
43 |
+
fig = plot_img_with_rects(image.permute(1, 2, 0).numpy(), pred, threshold, coef=192)
|
44 |
+
fig.savefig('figure_name1.png')
|
45 |
+
image = Image.open('figure_name1.png')
|
46 |
+
st.image(image)
|